jc138691 / kaggle-humpback

Code for 3rd place solution in Kaggle Humpback Whale Identification Challenge.

kaggle-humpback-submission

Code for 3rd place solution in Kaggle Humpback Whale Identification Challange.

To read the detailed solution, please, refer to the Kaggle post

Hardware

The following specs were used to create the original solution.

• Ubuntu 16.04 LTS
• Intel(R) Core(TM) i7-8700 CPU @ 3.20GHz
• 2x NVIDIA 1080 Ti

Reproducing Submission

To reproduce my submission without retraining, do the following steps:

1. Installation
2. Download Dataset
3. Download Pretrained models
4. Inference
5. Make Submission

Installation

All requirements should be detailed in requirements.txt. Using Anaconda is strongly recommended.

conda create -n humpback python=3.6
source activate humpback
pip install -r requirements.txt

Download dataset

Download and extract train.zip and test.zip to data directory. If the Kaggle API is installed, run following command.

$ kaggle competitions download -c humpback-whale-identification -f train.zip
$ kaggle competitions download -c humpback-whale-identification -f test.zip
$ unzip train.zip -d data/train
$ unzip test.zip -d data/test

Generate CSV files

You can skip this step. All CSV files are prepared in the data directory.

List of CSV files

filename	description
landmark.{split}.{fold}.csv	predicted landmarks for the train and test set
duplcate_ids.csv	list of duplicate identities
leaks.csv	leaks from post
split.keypoint.{fold}.csv	labels for training bounding box and landmark detector
train.v2.csv	label file that duplicate ids are grouped to single identity and several new whales are also grouped.

Landmark

To inference landmarks, run following commands

$ sh inference_landmarks.sh

Training

In the configs directory, you can find configurations I used to train my final models.

Train models

To train models, run following commands.

$ python train.py --config={config_path}

The expected training times are:

Model	GPUs	Image size	Training Epochs	Training Time
densenet121	1x 1080 Ti	320	300	60 hours

Average weights

To average weights, run following commands.

$ python swa.py --config={config_path}

The averages weights will be located in train_logs/{train_dir}/checkpoint.

Pretrained models

You can download pretrained model that used for my submission from link. Or run following command.

$ wget https://www.dropbox.com/s/fdnh29pjk8rpxgs/train_logs.zip
$ tar xzvf train_logs.tar.gz

Unzip them into train_logs then you can see the following structure:

results
  +- densenet121.1st
  |  +- checkpoint
  +- densenet121.2nd
  |  +- checkpoint
  +- densenet121.3rd
  |  +- checkpoint
  +- landmark.0
  |  +- checkpoint
  +- landmark.1
  |  +- checkpoint
  +- landmark.2
  |  +- checkpoint
  +- landmark.3
  |  +- checkpoint
  +- landmark.4
  |  +- checkpoint

Inference

If trained weights are prepared, you can create files that contain cosine similarities of images with target whales.

$ python inference.py \
  --config={config_filepath} \
  --tta_landmark={0 or 1} \
  --tta_flip={0 or 1} \
  --output={output_filepath}

To make submission, you must inference test and test_val splits. For example:

$ python make_submission.py \
  --input_path={comma seperated list of similarity file paths} \
  --output_path={submission_file_path}

To inference all models and make submission using pretrained models, simply run sh inference.sh

Post Processing

As you know, there are some duplicate whale ids. For the duplicate ids, the following process are applied.

Assume that the identity A and the identity B are duplicate.

1. If top1 prediction is the identity A, then I set the identity B to top2 prediction.
2. If the size of test image is equal to one of images in identity A and is not equal to any of images in identity B, then I set top1 prediction to identity A.